OT Score: Official Implementation

Official PyTorch implementation for the paper:

OT Score: An OT based Confidence Score for Prototype-Assisted Source Free Unsupervised Domain Adaptation

This repository contains code for source-free unsupervised domain adaptation (SFUDA) with the proposed OT score, including source model training, target adaptation, and OT-score-related feature/prototype utilities.

Overview

The OT score is a confidence score based on semi-discrete optimal transport. It is designed to estimate the reliability of target pseudo-labels in prototype-assisted source-free unsupervised domain adaptation.

This repository provides:

• train_source.py: train the source model and save source checkpoints
• train_target.py: adapt the target model from a trained source checkpoint
• utils/ot_score_utils.py: OT score computation and feature/prototype utilities

Repository Structure

.
├── train_source.py
├── train_target.py
├── environment.yaml
├── utils/
│   └── ot_score_utils.py
├── output/
└── README.md

1. Environment Setup

Create the conda environment from environment.yaml:

conda env create -f environment.yaml
conda activate OTscore

If you use a different environment name, for example DA, activate that environment consistently:

conda activate DA

2. Dataset Preparation

This code assumes that --datadir contains both:

1. Image folders referenced by the list files
2. DomainNet list files, such as:

clipart_list.txt
real_list.txt
painting_list.txt
sketch_list.txt

Each list file should follow the format:

relative/or/absolute/image_path label_id

For example:

clipart/airplane/image_0001.jpg 0
clipart/bicycle/image_0002.jpg 1

In the commands below, replace:

/path/to/domainnet/

with your own dataset root directory.

3. Train Source Model

To train a source model for one source-target task on DomainNet, run:

python train_source.py \
  --dset domainnet \
  --s 0 \
  --t 1 \
  --datadir "/path/to/domainnet/" \
  --output output

By default, this only trains the source model and does not run target-side testing.

To additionally run source-side testing after training, use:

python train_source.py \
  --dset domainnet \
  --s 0 \
  --t 1 \
  --datadir "/path/to/domainnet/" \
  --output output \
  --test_target

Source checkpoints will be saved to:

output/<dset>/<SourceInitial>/
  source_F_<seed>.pt
  source_B_<seed>.pt
  source_C_<seed>.pt

For example:

output/domainnet/C/
  source_F_2026.pt
  source_B_2026.pt
  source_C_2026.pt

4. Target Adaptation

After training the source model, run target adaptation with:

python train_target.py \
  --dset domainnet \
  --s 0 \
  --t 1 \
  --datadir "/path/to/domainnet/" \
  --output_src output \
  --output output \
  --seed 2026

The target adaptation script loads source checkpoints from:

<output_src>/<dset>/<SourceInitial>/source_*.pt

Target checkpoints will be saved to:

<output>/<dset>/<SourceInitial><TargetInitial>/
  target_F_<seed>.pt
  target_B_<seed>.pt
  target_C_<seed>.pt

For example:

output/domainnet/CR/
  target_F_2026.pt
  target_B_2026.pt
  target_C_2026.pt

5. Example Workflow

A typical workflow is:

Step 1: Train the source model

python train_source.py \
  --dset domainnet \
  --s 0 \
  --t 1 \
  --datadir "/path/to/domainnet/" \
  --output output \
  --seed 2026

Step 2: Adapt the target model

python train_target.py \
  --dset domainnet \
  --s 0 \
  --t 1 \
  --datadir "/path/to/domainnet/" \
  --output_src output \
  --output output \
  --seed 2026

6. OT Score Utilities

The OT score and related feature/prototype utilities are implemented in:

utils/ot_score_utils.py

These utilities are used during target adaptation to compute confidence scores for pseudo-labeled target samples.

The OT score can be used for:

• identifying low-confidence target pseudo-labels
• reweighting target samples during adaptation
• providing a label-free proxy for target-domain performance

7. Reference

If you find this repository useful, please cite our paper:

@article{zhang2026otscore,
  title={OT Score: An OT based Confidence Score for Prototype-Assisted Source Free Unsupervised Domain Adaptation},
  author={Zhang, Yiming and Liu, Sitong and Cloninger, Alex},
  journal={Transactions on Machine Learning Research},
  year={2026}
}